KR20030076661A - Method, module, device and server for voice recognition - Google Patents

Abstract

The present invention is a speech recognition method implemented in at least one terminal 114, the speech recognition method using a language model 311,

Detecting (502) at least one unrecognized representation in one terminal,

Recording (503) data representing the unrecognized representation (309) to the terminal;

Sending (603) the recorded data by the terminal to the remote server 116,

Analyzing (803) the data at the level of the remote server and generating (805) information for correcting the language model in view of at least one portion of the unrecognized representation;

Transmitting (806) said correction information from said server to at least one terminal (114, 117, 118) in order to be able to subsequently recognize at least a particular representation of said unrecognized representation.

It relates to a speech recognition method comprising a.

The invention also relates to a corresponding module, device 102, and remote server 116.

Description

Method, module, device and server for speech recognition {METHOD, MODULE, DEVICE AND SERVER FOR VOICE RECOGNITION}

The information or control system further increases the use of the voice interface to enable faster and / or more intuitive interaction with the user. As these systems become much more complex, the requirements in terms of speech recognition must be considered even more in terms of both recognition range (very much vocabulary) and recognition speed (real time recognition).

Speech recognition methods based on language models (probability that a given word of an application vocabulary is next to another word or group of words in the temporal order of the sentence) and the use of speech units are known in the art. These techniques are described in particular in the paper "Statistical methods for speech recognition" by Frederik Jelinek, published in 1997 by the MIT Press.

These techniques rely on speech models and speech units generated from representative speech samples (e.g., samples from a population of users of a terminal entering a command).

Indeed, the language model identifies the speaking styles commonly used by users of the system, and especially their "defects": hesitation, false start, change of mind, etc. Should be considered

The quality of widely used language models greatly affects the reliability of speech recognition. This quality is often measured by an index called the language model's perplexity, which roughly represents the number of choices the system must make for each decoded word. The lower this confusion, the better the quality.

The language model is needed to convert the speech signal into a textual string of words, which is often used by a dialog system. Then, to answer the question, it is necessary to construct the comprehension logic that enables the question to be understood.

There are two standard ways to create language models of many vocabularies:

The so-called N-gram statistical method, most often using bigrams or trigrams, includes N words where the probability of occurrence of a word in a sentence precedes that word, regardless of the rest of the context of the sentence. It is assumed to depend only on.

If we have an example of a trigram for a vocabulary of 1000 words, then we need to define a probability of 1000 ³ to define the language model, which is not possible. Therefore, these words are grouped into sets that are either explicitly defined by the model designer or inferred by a self-organizing method.

Thus, this language model is automatically constructed from textual language material.

This type of language model is mainly used in speech command systems, whose ultimate function is to convert speech signals into text without the need for any level of understanding.

The second method is a context-free grammer defined by a set of probable grammars, typically a so-called Backus Naur form, or BNF (Backus Naur Form), or extending this form to a contextual grammar. A grammar describes a syntax. Most rules describing grammar are written by hand. This type of language model is suitable for command and control applications followed by controlling the appliance after a speech recognition step or retrieving information in a database.

The language model of the application describes a set of expressions (eg sentences) that require the application to recognize. A disadvantage of the prior art is that, if the quality of the language model is poor, even if the recognition system performs very well at the acoustic-phonetic decoding level, the speech system has mediocre performance for certain expressions. Is to have.

This stochastic type of language model, when appropriately said, does not have a clear definition of what is in the language model and what is outside. Certain expressions only have a higher probability of occurrence a priori than other expressions.

Probable grammatical type language models show a clear difference between expressions belonging to a language model and expressions outside the language model. Thus, in these models, there is an expression that cannot be recognized at all, regardless of the quality of the speech model being used. These are generally expressions that have a meaning outside the field of application of the developed system or have no meaning at all.

Probabilistic types of language models and their derivative models have been found to be more effective for command and control applications. These grammars are often written by hand, and one of the major difficulties in developing dialog systems is to provide a high quality language model.

In particular, as far as the grammar type model is concerned, it may not be possible to specifically limit the language, especially if the language is easy to use by a large number of groups (e.g. in the case of remote controls for mass marketed appliances). . It may not be possible to take into account all possible expressions and turns of phrases (from formal language to slang) and / or errors in grammar.

The present invention relates to the field of voice interfaces.

More specifically, the present invention relates to the optimization of a language model and / or phonetic unit in a terminal using speech recognition.

1 is an overall schematic diagram of a system including a voice control box in which the techniques of the present invention may be implemented;

2 is a schematic diagram of a speech recognition box of the system of FIG.

3 is an electronic diagram of a speech recognition box implementing the schematic diagram of FIG.

4 is a schematic diagram of a server of the system of FIG. 1;

5 is a flow chart of a method for testing a representation by recording data relating to an unrecognized representation, implemented with the recognition engine of FIG.

6 is a flow chart of a method for transmitting data regarding an unrecognized representation, implemented by the reject module of FIG.

7 is a flow diagram of a method for receiving correction data, implemented by a module for loading the language model of FIG.

8 is a flow diagram of a method for receiving and processing correction data, implemented in the remote server of FIG.

The present invention relates to a speech recognition method and system that enables remote modification and improvement of a language model based on the recording of an expression not recognized by the system.

More specifically, a subject of the present invention is a speech recognition method implemented in at least one terminal, the speech recognition method using a language model,

Detecting at least one unrecognized representation in one terminal,

Recording at the terminal data representing the unrecognized representation;

Transmitting, by a terminal, the recorded data to a remote server over a first transmission channel;

At the level of the remote server, analyzing the data and generating information for correcting the language model taking into account at least one part of the unrecognized representation,

Transmitting the correction information from the server to at least one terminal via a second transmission channel to subsequently recognize at least a particular representation of the unrecognized representation.

In a voice recognition method comprising a.

Thus, the present invention updates several elements that enable the speech recognition as a function of a locally unrecognized representation, with a remote server having significant resources (e.g., human and / or computing power) for generating correction information. It relies on an entirely new and progressive approach to speech recognition.

This language model is

A strict sense language model (for example, when the data subject to recognition is purely text type);

A model formed of one or more language models of strict meaning and one or more sets of speech units (this is especially the case in speech recognition applied to speech samples)

Characterized in that it comprises a.

The present invention goes beyond a simple update of the vocabulary. Specifically, even if all words having a feature of the expression in the vocabulary have been used by the language model of the terminal, this expression may not be recognized. Only an update of the language model itself can allow this expression to be recognized later. An update of the vocabulary, one of the information items from which the language model is derived, is not sufficient.

Here, the expression is taken in a broad sense and relates to any verbal expression that enables interaction between the terminal and its user. Expressions (or words) include, in particular, sentences, phrases, isolated or non-isolated words, code words assigned to a terminal, instructions, commands, and the like. do.

The correction information may specifically include information that enables partial or complete modification of the speech unit and / or language model present at each terminal by deleting, replacing, or adding elements within the terminal.

The server may receive data from each terminal, thereby enabling each terminal to improve the speech unit and / or language model present at the data transmitting terminal and also at all other terminals, where each terminal is a server from all terminals. Benefit from the shared experience gained by

In this way, the present invention relates to a language style or phrase that is assigned to a particular user, for example, an expression other than "8pm" or "evening 8 o'clock", that is, an expression that was not provided during the construction of the implemented language model. "(A term difficult to imagine a priori) may be taken into account.

Furthermore, the present invention also takes into account the evolution of living languages (such as new phrases or expressions).

It is noted that the present invention applies equally well to probabilistic grammatical and stochastic type language models. When the present invention is applied to a probabilistic type language model, there is generally a great deal of correction data that affects recognition, whereas for probabilistic grammar type models there may be little correction data and also the effect and reliability of recognition. Can have a significant impact on

According to a particular characteristic, the method is characterized in that the data representing the unrecognized representation comprises compressed speech recordings representing parameters describing the acoustic signal.

Thus, the present invention allows to advantageously consider the voice data sent to the source for sophisticated analysis at the server level while limiting the amount of data sent to the remote server.

According to a particular characteristic, the method, during the step sent by the terminal, the terminal,

Contextual information on the use of speech recognition methods when no expression is recognized,

-Information about the speaker who spoke the unrecognized expression

And at least one item of information forming a part of the group including a further to the server.

Thus, speech recognition of expressions not recognized by the terminal that can be performed remotely is facilitated.

Furthermore, a check of the validity of the content of the unrecognized representation may be performed as a function of the context (eg, a "transmit record" command means meaning when the terminal to which this command was sent is a video recorder. And therefore valid but not meaningful for mobile phones).

According to certain features, the method is characterized by implementing scrambling of the encrypted and / or recorded data and / or correction information.

Thus, data is effectively made secure and confidential.

This information also relates to speech recognition modules using language models,

An analyzer for detecting unrecognized expressions,

A recorder for recording data representing at least one unrecognized representation,

A transmitter for transmitting the recorded data to a remote server;

A receiver for receiving correction information enabling correction of a language model transmitted to a correction module, so as to enable subsequent recognition of at least a particular expression among the expressions unrecognized by the module, the data being analyzed at the level of the remote server; A receiver, after generating information for correcting a language model in view of at least one portion of the later and unrecognized representation, the correction information being transmitted by a remote server

It relates to a speech recognition module comprising a.

The invention also relates to a speech recognition device using a language model,

An analyzer for detecting unrecognized expressions,

A recorder for recording data representing at least one unrecognized representation,

A transmitter for transmitting the recorded data to a remote server;

A receiver for receiving correction information enabling correction of a language model sent to the device to enable later recognition of at least a particular expression of the unrecognized expression by the device, after analyzing the data at the level of the remote server and A receiver in which the correction information is transmitted by a remote server after generating information for correcting the language model in view of at least one portion of the unrecognized expression

It relates to a voice recognition device comprising a.

The invention also relates to a speech recognition server in which recognition is implemented in a set of at least one remote terminal using a language model, comprising:

A receiver which forms part of a set of at least one remote terminal and receives data indicative of at least one representation not recognized by at least one terminal which has detected an unrecognized representation during a speech recognition operation,

A transmitter for transmitting to the set of at least one remote terminal the correction information obtained based on the analysis of the data received at the level of the server, the correction information being at least one of an unrecognized representation by each terminal of the set; Transmitter that allows correction of the language model to enable later recognition of the part

It relates to a speech recognition server, characterized in that it comprises a.

Since the specific features and benefits of the speech recognition module, device, and server are similar to those of the speech recognition method, their features and benefits are not described herein again.

Other features and advantages of the present invention will become more apparent upon reading the following detailed description of the preferred embodiments and the accompanying drawings, which are given by way of non-limiting illustrative examples only.

The general principle of the present invention therefore depends on the speech recognition implemented at the terminal, using a set of speech units and / or language models that can be updated by the remote server, especially when the speech recognition method is deemed necessary. do.

In general, each terminal may recognize a representation (eg, a sentence or a command) configured by the speaker and perform a corresponding action.

Nevertheless, it is often found that certain expressions that are fully understandable to a human are not recognized by a device or module that implements speech recognition.

Unrecognized reasons include a number of reasons,

Vocabulary used by speakers who do not form part of the language model,

-Specific pronunciation (e.g. with accents),

A specific phrase turn not provided by a speech recognition device or module,

- Etc

This can be.

Specifically, the language model and the set of speech units take into account a sample of expressions customarily used by a typical group, ie words, pronunciation, and / or phrases of phrases not previously considered (and not considered). Often constructed based on statistical data.

The present invention relies on first detecting the expression not recognized by the speech recognition device or module.

When a representation is not recognized, the terminal records data representing a signal corresponding to the unrecognized representation in order to send the unrecognized representation to the remote server (e.g., voice digital recording of that representation). .

At the level of the remote server that centralizes the unrecognized representation from the set of terminals, a human operator can analyze this unrecognized representation.

Certain of these unrecognized expressions may be found to be incomprehensible and / or unusable and will then be discarded.

On the other hand, other expressions would be fully understood by an operator who could "translate" those expressions thus far unrecognized by the terminal through human / machine connections (if the operator finds them useful) to code understandable at the server. Could be.

The server may take into account these expressions and their translations to generate information for correcting the language model and / or speech unit set.

The correction here is

Modification of the model and / or

Complement the model

It should be noted that it is understood as.

The server then sends this correction information to each terminal, where each terminal can update its language model and / or speech unit set, filled with multiple representations not recognized by itself or by another terminal.

Thus, speech recognition of each terminal is improved by benefiting from the experience shared by all terminals.

According to a particular mode of the present invention, the analysis is not performed by an operator but by a server that can have much more resources at random than a simple terminal.

According to a particular embodiment, the terminal may include server context data (e.g., control, location, terminal performed manually or verbally after a failure of a time, date, or voice command) with data representing a signal corresponding to an unrecognized representation. Type, etc.).

This may facilitate the analysis of the operator and / or server.

An overall schematic diagram of a system including a voice control box in which the techniques of the present invention may be implemented is depicted in connection with FIG. 1.

This system,

A remote server 116 controlled by an operator 122 which is a human,

Multiple user systems 114, 117 and 118

In particular it includes.

Remote server 116 is linked to respective user systems 114, 117, and 118 via communication downlinks 115, 119, and 120, respectively. These links may be permanent or temporary and may be of any type well known to those skilled in the art. These links may be of particular broadcast type and may be based on the RF, satellite or wired channel used by the TV or any other type, for example an internet type link.

1 specifically describes a user system 114 linked to a server 116 via a communication uplink 121. This link can likewise be of any type well known to those skilled in the art (especially telephone, internet, etc.).

User system 114

A voice source 100, which may in particular be configured as a microphone for picking up a voice signal generated by the speaker,

A speech recognition box 102,

A control box 105 for driving the appliance 107,

Controlled appliance 107, for example of TV, video recorder or mobile communication terminal type

A unit 109 for storing the detected expression as unrecognized,

An interface 112 that enables upward and downward communication to the server 116.

In particular, it includes.

The source 100 is linked to the speech recognition box 102 via a link 101, which allows the source 100 to transmit sound waves of an analog source representing the speech signal to the box 102. do.

Box 102 retrieves contextual information 104 (eg, a list of control codes or information such as the type of appliance 107 that can be controlled by control box 105) via link 104. A command can be sent to the control box 105 via the link 103.

The control box 105 is an information function that the box 102 recognizes according to its language model and dictionary, for example, sending a command to the appliance 107 via a link 106 such as infrared. do.

Control box 105 detects representations that box 102 does not recognize and transmits unrecognized signals instead of simply rejecting these representations, thereby allowing box 102 to store these representations via link 108. Recording to 109 is performed.

The unit 109 for storing the unrecognized representation transmits representative data to the interface 112 via the link 111, which interface 112 relays this data to the server 116 via the link 121. do. After the correct transmission, the interface 112 may transmit a signal 110 to the storage unit 109, which may then erase the transmitted data.

The control box 105 further receives correction data from the interface 112 via the link 113 that the interface 112 receives from the remote server via the link 115. These correction data are considered by the control box 105 for updating the set of speech units and / or language model.

According to the contemplated embodiment, the source 100, the speech recognition box 102, the control box 105, the storage unit 109, and the interface 112 are one and form part of the same device, thus linking ( 101, 103, 104, 108, 111, 110 and 113 are links inside this device. This link 106 is typically a wireless link.

In accordance with a first variant of the invention described in FIG. 1, the elements 100, 102, 105, 109, and 112 are partially or completely separated and are one and do not form part of the same device. In this case, the links 101, 103, 104, 108, 111, 110, and 113 are external wires or other links.

According to a second variant embodiment, the source 100, the boxes 102 and 105, the storage unit 109, and the interface 112 as well as the appliance 107 are one and form part of the same device, and have an internal bus. Mutually linked by {links 101, 103, 104, 108, 111, 110, 113, and 106}. This variant embodiment is particularly advantageous when the device is for example a mobile phone or a portable communication terminal.

FIG. 2 shows a schematic diagram of a box that is voice controlled such as the box 102 shown for FIG. 1.

The box 102 receives sound waves 101 of analog sources from the outside, which are processed by an acoustic-phone decoder 200 (Acoustico-Phonetic Decoder) or APD (also called "front end"). It is to be noted. The APD 200 generates a real vector, or vector, belonging to a code book, typically representing an oral resonance, transmitted to the recognition engine 203 via a link 201. For this purpose, the source sound waves 101 are sampled at regular intervals (typically every 10 ms). APD is, for example, authored by Heynek Hermansky and published in the Journal of the Acoustical Society of America Vol. 97, No4, 1990, based on PLP ("Perceptual Linear Prediction"), which is described in particular in the paper "Perceptual Linear Prediction (PLP) analysis of speech" published in pages 1738-1752. I put it.

With the help of the dictionary 202, the recognition engine 203 uses a hidden Markov model, in particular an HMM and a language model (a model representing the probability of one word coming out according to another word). 203 analyzes the received real vector. The recognition engine is described in detail in the book "Statistical Methods for Speech Recognition," published by Frederick Jelinek and published in 1997 by MIT Publishing.

The linguistic model allows the recognition engine 203 (especially the use of hidden Markov networks) to determine which words can follow a given word of any expression available by the speaker in a given application and determine the relevant probabilities. Can provide. Relevant words belong to the vocabulary of an application, which may be small (typically 10 to 300 words) or large (e.g., larger than 300 000 words) regardless of the language model.

The patent application PCT / FR00 / 03329 filed November 29, 1999, filed under the name of Thomson Multimedia, describes a language model comprising a plurality of syntactic blocks. The use of the present invention, which is the subject of this patent application, is particularly advantageous in connection with this type of module language model, as the module can be updated independently to avoid downloading an excessively large amount of files.

The language model is sent by the language model loading module 207. This module 207 receives a language model transmitted from the server via link 113 and receives update or correction data of the voice unit and / or language model transmitted from the server.

It is noted that the dictionary 202 belongs to a language model with reference to words from this dictionary. Thus, this dictionary 202 may be updated and / or corrected via the language model loaded by module 207.

After implementing the recognition operation based on the use of the Viterbi algorithm, the recognition engine 203 supplies to the reject module 211 an ordered list of word strings according to the language model representing the best score for the spoken expression. do.

The reject module 211 operates downstream of the recognition engine 203 and operates according to one of the following principles:

Sometimes, for reasons specific to the Viterbi algorithm, the Viterbi algorithm may not produce a consistent list because its score is so low that it exceeds the limits of the machine's acceptable accuracy in terms of arithmetic operations. Therefore, no coherent perfection exists. Thus, this representation is rejected when the reject module 211 detects one or more scores below a predetermined acceptable limit.

Each element of the list computed by the Viterbi algorithm has been retained because the associated score was among the best relative scores of all possible representations according to the language model. In addition, the Markov network associated with each of these representations may be used to evaluate the intrinsic probability of the associated network generating the representation associated with the observed score. This rejection module 211 analyzes this probability and if this probability is less than a predetermined threshold of acceptable probabilities, this representation is rejected.

According to another method, the reject module 211 performs the complementary processing of this representation using criteria not considered in the Viterbi development process for the best proposal obtained through the Viterbi algorithm. For example, the reject module checks whether these signal portions that should be uttered are actually such because the signal portions that should be uttered are associated with a vowel. If the proposed expression does not meet these conditions, this expression is rejected.

If the reject module 211 rejects any expression as illustrated above, the expression is said to be unrecognized and a signal indicative of the rejected expression is sent to the recognition engine 203. At the same time, the reject module transmits the recording of the unrecognized representation to the storage unit 109 via the link 108.

Recognition engine 203 is responsible for recognizing representations coming from APD 200 in the form of speech samples. Thus, the recognition engine 203

A speech unit for constructing a phonetic representation of the word in the form of a Markov model, in which each word of the dictionary 202 may possibly have several "phonetizations"

A strict semantic language model for recognizing larger or lower complexity representations

Use

The recognition engine 203 understands these representations by the appliance 107 to understand the representations (that is, representations not rejected by the module 211) and the representations that the recognition engine 203 has identified based on the received vectors. To means 205 for translating into instructions that can be made. This means 205 employs an artificial intelligence translation process that takes into account the contextual information 104 supplied by the control box 105 before sending one or more instructions 103 to the control box 105. use.

FIG. 3 schematically illustrates a speech recognition module or device 102 implementing the schematic diagram of FIG. 2 as shown in conjunction with FIG. 1.

Box I02 is interconnected by an address and data bus,

A voice interface 301,

An analog-to-digital converter 302,

A processor 304,

Non-volatile memory 305,

A random access memory 306,

A receiving module 312,

A sending module 313,

Input / Output Interface (307)

It includes.

Each element shown in FIG. 3 is well known to those skilled in the art. These ordinary elements are not described here.

The word "register", which is used throughout the description, is both a small memory area (several data bits) and a large memory area (an area capable of storing entire rows of entire program or transaction data) in each memory mentioned. To indicate.

Nonvolatile memory 305 (ROM) specifically holds a program for operating processor 304 in " prog "

The random access memory 306 retains data, variables, and intermediate processing results in registers having the same name as the data held by the registers for convenience.

A register 309 ( Exp_Not_Rec ) in which the recording of the unrecognized representation is kept,

A counter 310 ( Nb_Exp_Not_Rec ) of the unrecognized sentence,

Language model ( Model_Language ) in register 311

In particular, it includes.

Receive module 312 and transmit module 313 are modules that enable data to be transmitted from or to remote server 116, respectively. Wired or wireless technologies for receiving and transmitting are well known to those skilled in the telecommunications art and are not described in greater detail.

4 illustrates a server 116 of the system shown with respect to FIG. 1.

The server 116 is controlled by the operator 122 who is a person via any human / machine interface 404 (eg, a keyboard and screen type interface).

Server 116

A receiver 400,

An analyzer 401,

A module 402 for configuring correction data of the speech unit set and / or the language model,

Transmitter 403

In particular, it includes.

Receiver 400 is compatible with transmitter 313 of the terminal and each terminal has specific data (e.g., recording) that represents an unrecognized representation and possibly complementary data (e.g. contextual data) via link 121. Can be received from.

The analyzer 401 receives a data set coming from the receiver 400 via a link 121, which transmits the data set to the operator 122 via the interface 404. 122), for example,

Is equipped with a screen and a keyboard enabling interaction with the server 116 and its controls,

-Equipped with a speaker or audio headset for listening to unrecognized recordings

It is a terminal.

The interface 404 also allows the analyzer 401 to:

Whether unrecognized representations not covered by the language model remain incomprehensible, have no meaning in the application for the terminal, and / or are not related to the terminal (and thus this should not be included in the language model) In this case, this expression is later ignored for correction of the language model and discarded by the analyzer 401,

Nevertheless, whether the unrecognized expressions belong to a strict meaning language model (this involves a pure recognition problem), and in this case it involves modification of the speech unit, not a strict meaning language model,

Which representation of the content has been identified by the operator and then translated into a control code, for example, where the unrecognized representation does not belong to a language model and does not have meaning for the intended terminal. This entails correcting a strict language model,

Information from the operator 122 can be received.

It is also possible to combine the second and third solutions, in which case this involves modifying both the strict meaning language model and the speech unit.

This embodiment corresponds to manual processing of unrecognized expressions. According to this embodiment, the operator 122 who is a human listens to the unrecognized expression and analyzes the cause of the rejection reason. The operator 122 determines in particular whether or not this representation belongs to a language model. If this expression belongs to a language model, the operator uses the expression (which belongs to a language model that should have been recognized but not recognized by other reasons, such as noise, speaker accents, etc.) to identify an intrinsic recognition problem. Analyze

According to the first variant embodiment, the process is automatic and eliminates human intervention. In this case, the server 116 and especially the analyzer 401 possess a relatively significant computing power which can be much larger than the terminal in particular. According to this variant embodiment, the analyzer 401 analyzes each unrecognized expression in a more appropriate way than can be done by the terminal, for example using a better language model and / or a more complex speech model. do. Without receiving this rigorous real-time computational requirement, which may be a terminal (which often requires a fast response time to the speaker's command), the analyzer 401 may also allow for recognition that requires a longer processing time than, for example, at the terminal. Can be.

According to the second variant embodiment, the process is semi-automatic and the human intervention of the operator is limited to cases which cannot be solved by the analyzer.

The general structure of the server 116 is described here similarly to that of the terminal as described for FIG. 3 according to a preferred embodiment, and in particular interconnected by an address and data bus,

A processor;

Random access memory,

Non-volatile memory,

An appropriate transmission module,

A receiving module,

-Human / machine connection interface

It includes.

In accordance with FIG. 5, which illustrates a flowchart of the recording of data relating to an unrecognized representation, which is implemented by the recognition engine 203 of FIG. 2, during the first initialization step 500, the microprocessor 304 is The execution of the program 308 is started and the variables of the random access memory 306 are initialized.

Thereafter, during the presentation wait step 501, the microprocessor waits and receives the presentation sent by the speaker.

Then, during the test 502, after performing a speech recognition operation on the received representation, the microprocessor recognizes the representation according to one or more criteria shown with respect to the description of the reject module 211 of FIG. 2. Determine whether it is recognized or not recognized.

If so, during the control step 504, the speech recognition box 102 considers the result of the speech recognition applied to the received representation and performs the appropriate action, for example a command.

If not recognized, during the step 503 of recording the representation, the unrecognized representation is compressed and recorded in the storage unit 109 waiting for transmission to the remote server 116 shown with respect to FIG.

When either step 503 or 504 is completed, step 501 awaiting the presentation is repeated.

FIG. 6 shows a flow diagram of the transmission of data relating to an unrecognized representation, implemented by the reject module of FIG. 2, during the first initialization phase 600, the microprocessor 304 starts executing the program 308. The variable of the random access memory 306 is initialized.

The microprocessor 304 then waits for the recording of the unrecognized representation and then receives it during step 601 of waiting for the unrecognized representation by the speech recognition box 102.

Then, during step 602, terminal 114 connects to remote server 116 according to methods well known to those skilled in the art of telecommunications.

Then, during step 603, recording of the unrecognized representation is made and transmitted to the remote server 116.

Then, during disconnection step 604, the terminal is disconnected from remote server 116, and the signal is for storing data corresponding to the unrecognized representation representing the transmission of the recording of the interface 112 and the representation with the remote server. Is transmitted between units 109. At this time, the data corresponding to these representations are erased from the storage unit 109.

Then step 601 is repeated.

FIG. 7 shows a flow diagram of receiving correction data, implemented by module 207 for loading the language model of FIG. 2.

After the first initialization step 700, during step 701, the terminal is placed in a waiting state waiting for correction data broadcast by the server 116 to the plurality of terminals.

Then, during step 702, the terminal considers the correction data to update the speech unit set and / or language model used by the speech recognition module. Depending on the nature of the correction data, these data

Replacing existing data in the speech unit set and / or language model,

Modifying existing data,

Complementing existing data, and / or

Performing deletion of existing data

can do.

After performing step 702, step 701 is repeated.

8 shows a flow diagram for receiving and processing correction data, implemented within the remote server of FIG. 4.

After the first step 800 of initializing the induction and parameters of the program for the management of the server, the server 116 receives from the terminal (terminal performing the step 602 shown in FIG. 6). It is placed in a waiting state waiting for an originating connection request and establishes a connection with the terminal according to methods well known to those skilled in the telecommunications arts.

Subsequently during step 802, server 116 receives data originating from the connected terminal performing step 603 described above. These data include, in particular, the recording of one or more representations rejected by the terminal due to not being recognized by the speech recognition module implemented in the terminal. When all data has been received, the connection between the terminal and the server 116 is broken.

Then, during processing 803 of the received data, the server 116 records the representations received automatically or semi-automatically by the operator 122 or in accordance with various alternatives shown for FIG. 4. Process each.

Then, during the test 804, the server 116 determines in particular whether one or more representations received were understandable and relevant to the representation or the terminal that sent these representations. In this case, it is necessary to update the language model and / or the voice unit.

If not, the wait step 801 is repeated.

In the opposite case, server 116 constructs the correction data of the language model, which may take several forms to enable step 607 (illustrated above) in the terminal after receiving the correction data. These correction data are especially

-Indicators that specify the nature of the correction, in particular substitutions, modifications, supplements, or deletions,

Correction data as a function of the indicator

It includes.

It should be noted that if a language model includes a plurality of syntax blocks (especially for language models as described in the aforementioned patent PCT / FR00 / 03329), each module can be corrected individually. In this case, the correction data also includes an indicator of the module or modules to be corrected.

Thereafter, during step 806, the server 116 broadcasts the correction data to one or preferably a set of terminals that can update the set of speech units and / or its language model in accordance with step 607.

Then step 801 is repeated.

The procedure is repeated this way and can be repeated several times. This procedure also allows the application to be updated by adding new questions.

Of course, the invention is not limited to the exemplary embodiments described above.

In particular, one of ordinary skill in the art may change the definition of a terminal implementing the present invention relating to a module and / or any type of device that may or may use a speech recognition method (eg, a multimedia terminal). , Type of TV, video recorder, multimedia digital decoder (or set-top box), audio or video equipment, fixed or portable terminal.

Likewise. The invention relates to any type of remote server (e.g., an Internet server, equipment connected to a TV program station, equipment connected to a mobile communication network, service provider equipment, etc.)

Furthermore, according to the invention, the transmission channel for data corresponding to an unrecognized sentence and the transmission channel for data for correcting a language model and / or a speech unit are arbitrary and in particular

An RF transmission pathway,

A satellite transmission path,

-Channels of TV broadcast networks,

A channel of an internet type network,

-The channel of the telephone network,

-Channels on mobile networks,

Removable media

It includes.

Furthermore, the present invention not only relates to unrecognized sentences, but also to, for example, one or more sentences, isolated or unisolated words, phrases that enable communication between the machine and its users. It is to be noted that it relates to any type of verbal expression, such as voice code. These oral expressions are, for example, information data that allows a user to send configuration data, programming data, and the like to a machine, to form the subject of conversation between the machine and the user as well as instructions. It may be associated with any type of data that may be.

The method of updating the language model described in this patent supports not only strict speech recognition methods but also orthographic mistakes and / or typing mistakes, and also as described in this patent, the Markov model. It should also be noted that the same applies to a method of recognizing text input based on a Markovian model or a language model.

It will be appreciated that the invention is not limited to the installation of pure hardware but can also be implemented in the form of instruction strings of computer programs or in any form that mixes hardware and software portions. If the present invention is installed in part or in whole software form, the corresponding instruction string may be stored in removable storage means (e.g., a diskette, CD-ROM or DVD_ROM) or otherwise The storage means is partially or wholly readable by a computer or microprocessor.

As described above, the present invention can be used in the field of voice interface and the like.

Claims (7)

As a speech recognition method implemented in at least one terminal 114 and using a language model 311, Detecting (502) at least one unrecognized representation in one of said terminals, Recording (503) data representing the unrecognized representation (309) to the terminal; Transmitting (603) the recorded data by the terminal to a remote server 116 via a first transmission channel 121; Analyzing (803) the recorded data at the level of the remote server and generating (805) information for correcting the language model in view of at least one portion of the unrecognized representation; The correction information from the server to at least one terminal 114, 117, 118 via a second transmission channel 115, 119, 120 in order to be able to subsequently recognize at least a particular representation of the unrecognized representation. Sending 806 Characterized in that the voice recognition. 2. The method of claim 1, wherein the data representing the unrecognized representation (309) comprises compressed speech recordings representing parameters depicting an acoustic signal. The method of claim 1 or 2, wherein during the transmission step by the terminal, the terminal Contextual information on the use of the speech recognition method when no expression is recognized, -Information about the speaker who spoke the unrecognized expression And further transmitting to the server at least one item of information forming part of a group comprising a. 4. A method according to any one of the preceding claims, characterized in that it implements encryption and / or scrambling of said recorded data and / or said correction information. A speech recognition module 102 using a language model, An analyzer for detecting unrecognized expressions, A recorder for recording data representing at least one unrecognized representation, A transmitter for transmitting the recorded data to a remote server; A receiver for receiving correction information enabling correction of the language model sent to the module to enable subsequent recognition of at least a particular expression of the unrecognized expression by the module, the correction information being the data After analyzing at the level of the remote server and generating information for correcting the language model taking into account at least one portion of the unrecognized expression, the receiver being transmitted by the remote server Voice recognition module, characterized in that it comprises a. As speech recognition device 102 using a language model, An analyzer for detecting unrecognized expressions, A recorder for recording data representing at least one unrecognized representation, A transmitter for transmitting the recorded data to a remote server; A receiver for receiving correction information enabling correction of the language model transmitted to the device so as to enable later recognition of at least a particular expression of the unrecognized expression by the device, the correction information being the data After analyzing at the level of the remote server and generating information for correcting the language model taking into account at least one portion of the unrecognized expression, the receiver being transmitted by the remote server A speech recognition device, characterized in that it comprises a. A speech recognition server 116 in which speech recognition is implemented in a set of at least one remote terminal using a language model, A receiver which forms part of the terminal set and receives data indicative of the unrecognized at least one representation by at least one terminal which detected the unrecognized representation during a speech recognition operation; A transmitter for transmitting correction information obtained based on the analysis of the data received at the level of the server to the terminal set of at least one remote terminal, the correction information being recognized by each terminal of the terminal set; A transmitter that enables correction of the language model to enable later recognition of at least one portion of a given expression Characterized in that it comprises a, speech recognition server.